The Claude Code Prompting Playbook

Most prompting advice is a list of do’s and don’ts you can’t verify. The version Anthropic’s own engineers use is different: it starts from evals, treats a prompt like code you debug one failure at a time, and knows when the answer isn’t a better prompt at all. Here’s that playbook, the way I apply it to production LLM work.

Two situations cover almost everything: you’re maintaining a prompt that’s drifting (often after a model migration), or you’re building a new agent from scratch. The techniques overlap, and they’re the same ones behind the pipelines at Geonimo.

Start from evals, not opinions

You can’t tell whether a prompt change helped unless you measure it. So before touching the wording, build a small eval suite. It needs three kinds of cases: a control that should always pass (unambiguous, the model handles it easily), edge cases where you’ve seen the model fail before (so regressions can’t sneak back), and capability boundaries, where the right move is to hand off to a human or refuse outright. With that suite, every change becomes a measurable experiment instead of a vibe.

Clean up before you optimize

Most underperforming prompts improve just from hygiene. Real prompts accumulate cruft: instructions copied from a website, a line telling the bot it’s human, patches for three models ago, all mashed into one paragraph. Fix the structure first. Use XML tags to separate role, guidelines, policy, tone, and data. Remove the redundant text. The rule I keep coming back to: if you’re reading a prompt and can’t tell the guidelines from the policy from the data, the model can’t either. Then add an output contract, define the output format, and enforce it in the harness with a stop sequence or structured outputs when the schema is complex.

Target failure modes one at a time

With a clean prompt and a failing eval, fix one failure mode at a time. Three patterns come up constantly:

Building new: it’s prompt, model, and harness

From scratch, the prompt is only one of three levers, the model and the harness matter just as much. Hill-climb across all three. A real example: a scheduling agent failed every case on a small model with a simple prompt. A larger model reduced the errors but still failed. The larger model with adaptive thinking finally passed, but tripled tokens and latency. A smaller model with a much better prompt got partway. What won was going agentic: splitting the job into three simple prompts, a generator, an evaluator that reports specific violations, and a repairer that fixes them, which solved every case with lower cost and latency than brute force.

That generate-evaluate-repair loop has a bonus: you can inject soft constraints at runtime (“keep Harry and Sally on different shifts”) without rewriting the backend. Choosing among prompt, model, and harness is a cost / latency / quality decision, exactly the kind of trade-off I make daily on production pipelines, and the through-line of how I build.

This pairs with the habits Anthropic’s team uses and how to decide between tools, skills, and subagents. If you want help making your AI systems reliable in production, let’s talk.